.png){kind=icon}
.png){kind=icon}
Return
Debiasing Long-Tailed Hematopoietic Cell Detection Based on Category Distribution Alignment
.png){kind=icon}
PRE
AI
W
L
J
W
Y
DOI：10.1002/ima.70405.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Hematological morphological analysis is a central clinical method for revealing hematopoietic function and assessing disease risk, encompassing evaluation of both bone marrow and peripheral blood. The task is challenging because hematopoietic cells comprise many subtypes with only subtle inter-class differences and substantial intra-class variability. Moreover, many subtypes occur predominantly in patients with specific diseases, producing a pronounced long-tailed distribution that tends to bias classifiers toward common categories. Although prior work has begun to address hematopoietic cell detection, it has often neglected this long-tailed imbalance. Here, we propose a debiasing framework for long-tailed hematopoietic cell detection. We fit a Gaussian mixture model to the outputs of a hematopoietic cell detector to estimate the model's biased category distribution. Using ground-truth labels, we construct an unbiased, data-rich real-category distribution. We then align the detector's category distribution with this real distribution to recalibrate decision boundaries for each category. The proposed method also functions as an independent plug-in that can calibrate classification bias in diverse object detection models. We evaluated the approach on our private PBMCD dataset and on a combined set comprising PBMCD and the publicly available WBCDD dataset. The proposed method outperformed baseline models and produced especially marked gains on the tail category.
Keywords:
category distribution fine-tuning
cell detection
inter-category difference loss
long-tailed distribution
Journal
IF:
2.5
Papers:
2.1K
Citations:
2.3K
Organization
Cited Papers